ai-berkshire
Frameworkxbtlin/ai-berkshire
AI-era value investing research framework based on Claude Code with multi-agent adversarial analysis.
Overview
AI Berkshire is a structured skill set for Claude Code that systematizes the methodologies of four value investing masters (Buffett, Munger, Duan, Li). It enables deep, reproducible company research through multi-agent parallel analysis, cross-validated financial data, and bias-resistant checklists.
README Preview
中文 | [English](README_EN.md)\n\n# AI Berkshire - AI 时代的价值投资研究框架\n\n> "Price is what you pay, value is what you get." — Warren Buffett\n>\n> 用 AI 重新定义投资研究的深度与效率。\n\n**AI Berkshire** 是一套基于 [Claude Code](https://claude.ai/code) 的投资研究 Skill 合集，将巴菲特、芒格、段永平、李录四位价值投资大师的方法论系统化、结构化，通过 AI Agent 实现专业级投资研究。\n\n一个人 + Claude = 一个投研团队。\n\n---\n\n## Real Track Record\n\n> 不是纸上谈兵。这套框架背后是真金白银验证的投资体系。\n\n### 2024 全年收益：+69.29%\n\n\n\n### 2025 年至今收益：+66.38%\n\n\n\n### 与主要指数对比\n\n| 指标 | 2024 全年 | 2025 至今 |\n|------|----------|----------|\n| **本框架实盘** | **+69.29%** | **+66.38%** |\n| 恒生指数 | +17.67% | +27.77% |\n| 标普500 | +23.31% | +16.39% |\n| 沪深300 | +14.68% | +17.66% |\n| 纳斯达克 | +28.64% | +20.36% |\n\n**2024 年超额收益**：跑赢标普500 **46个百分点**，跑赢恒生指数 **52个百分点**\n\n**2025 年超额收益**：跑赢标普500 **50个百分点**，跑赢恒生指数 **39个百分点**\n\n**两年累计实盘收益超 146万元**，连续两年大幅跑赢全球主要指数。\n\n> *免责声明：历史收益不代表未来表现。截图来自富途证券真实账户。*\n\n---\n\n## 为什么不能直接问 AI？\n\n你当然可以直接问 Claude："帮我分析拼多多值不值得买"。你会得到一篇"一方面...另一方面..."的平衡分析，最后以"投资有风险，请自行判断"收尾。\n\n**这种分析看起来对，但没法拿来做决策。**\n\nAI Berkshire 解决的不是"能不能分析"的问题，而是**分析质量和决策纪律**的问题。以下是核心差异：\n\n### 1. 强制给结论，不打太极\n\n直接问AI，你得到的是两面讨好的"分析"。AI Berkshire 强制输出：**通过/不通过/灰色地带**，带具体价格区间和分层建议。\n\n> 普通AI回答：*"拼多多有增长潜力但也面临竞争压力，投资者需要权衡..."*\n>\n> AI Berkshire 输出：\n\n> | 策略 | 建议 | 价格区间 |\n> |------|------|---------|\n> | 激进型 | 当前价位可建仓20% | $95-105 |\n> | 稳健型 | 等回购政策明确后建仓 | $85-95 |\n> | 保守型 | 不符合10年确定性标准，观望 | — |\n>\n> **镜子测试**：5句话说不完整 = 不买，没有例外。\n\n### 2. 四大师视角对抗，而非单一分析\n\n不是"用巴菲特方法分析一下"这么简单。四个视角会产生**真实的矛盾和张力**——\n\n以拼多多为例：\n- **段永平**（商业模式）：好生意，C2M模式难以复制 → 评分 3.7/5\n- **巴菲特**（财务估值）：扣现金PE仅6.3x，印钞机 → 评分 4.4/5\n- **芒格**（逆向思考）：护城河比想象中浅，抖音3年做到4万亿GMV → 评分 3.5/5\n- **李录**（长期确定性）：管理层文化有隐患，10年后不确定 → 评分 2.0/5\n\n**巴菲特说"真便宜"，李录说"不确定就不买"**——这种冲突才是投资决策的真实状态。单一prompt无法制造这种多视角对抗，而这恰恰是避免盲点的关键。\n\n### 3. 结构化反偏见机制\n\nAI最危险的不是给错答案，而是给一个**看起来很对但经不起推敲**的答案。AI Berkshire 在流程中内置了多层"防骗"机制：\n\n| 机制 | 解决什么问题 | 举例 |\n|------|------------|------|\n| **信息丰富度评级（A/B/C）** | 防止"资料多=确定性高"的幻觉 | 泡泡玛特评为B级：数据有限，推算指标标注置信度 |\n| **芒格式逆向检验** | 强制思考失败场景 | "什么情况下拼多多
FAQ (2)
informationalWhat is Pony Ma's Tencent shareholding in 2025 and what percentage of Tencent's net profit is his RMB 42.92 million salary?
Use Tencent's 2025 annual report figure: Pony Ma holds 804,859,700 shares, equal to 8.82%. The previous 8.72% was the 2024 annual report figure; his share count did not change, but the ratio rose passively because repurchased shares were cancelled. For the salary ratio, calculate: RMB 42,920,000 ÷ RMB 259,600,000,000 (2025 Non-IFRS net profit) = 0.0001653, or 0.01653%. Rounded, this is 0.0165%, equivalent to 1.65 per 10,000 (万分之1.65). Replace the old wording "0.0017%" or "十万分之一点七"; both understate the ratio by 10 times.
TroubleshootingHow do I fix WebSearch being blocked by permissions for the research agents in the Claude Code investment-team skill?
Grant WebSearch permission and restart the workflow. In Claude Code, run /permissions, remove any WebSearch deny rule, and add WebSearch to the allow list. You can also edit ~/.claude/settings.json or the project .claude/settings.json:
{
"permissions": {
"allow": ["WebSearch"],
"deny": []
}
}
Restart Claude Code and run /investment-team 兴发集团 again, approving WebSearch for each research agent. Verify with a direct request such as "Use WebSearch to find 兴发集团 latest results". If it remains blocked, check your organization's managed policy; a policy-enforced deny cannot be overridden locally. If direct WebSearch works but the skill still blocks it, the investment-team subagent configuration must explicitly allow the WebSearch tool. As a temporary workaround, run the searches in the main Claude Code session and paste the latest filings or results into the workflow, since reports generated without WebSearch rely only on training data.